vad : add initial Voice Activity Detection (VAD) support (#3065)

* vad : add initial Voice Activity Detection (VAD) support

This commit add support for Voice Activity Detection (VAD). When enabled
this feature will process the audio input and detect speech segments.
This information is then used to reduce the number of samples that need
to be processed by whisper_full.

Resolves: https://github.com/ggml-org/whisper.cpp/issues/3003

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

src/whisper-arch.h

@@ -139,3 +139,59 @@ static const std::map<asr_tensor, ggml_op> ASR_TENSOR_INFO = {
     {ASR_TENSOR_ATTN_OUT_WEIGHT,       GGML_OP_MUL_MAT},
     {ASR_TENSOR_ATTN_OUT_BIAS,         GGML_OP_ADD},
 };
+
+enum vad_tensor {
+    VAD_TENSOR_STFT_BASIS,
+    VAD_TENSOR_ENC_0_WEIGHT,
+    VAD_TENSOR_ENC_0_BIAS,
+    VAD_TENSOR_ENC_1_WEIGHT,
+    VAD_TENSOR_ENC_1_BIAS,
+    VAD_TENSOR_ENC_2_WEIGHT,
+    VAD_TENSOR_ENC_2_BIAS,
+    VAD_TENSOR_ENC_3_WEIGHT,
+    VAD_TENSOR_ENC_3_BIAS,
+    VAD_TENSOR_LSTM_WEIGHT_IH,
+    VAD_TENSOR_LSTM_WEIGHT_HH,
+    VAD_TENSOR_LSTM_BIAS_IH,
+    VAD_TENSOR_LSTM_BIAS_HH,
+    VAD_TENSOR_FINAL_CONV_WEIGHT,
+    VAD_TENSOR_FINAL_CONV_BIAS,
+};
+
+static const std::map<vad_tensor, ggml_op> VAD_TENSOR_OPS = {
+    {VAD_TENSOR_STFT_BASIS,          GGML_OP_IM2COL},
+    {VAD_TENSOR_ENC_0_WEIGHT,        GGML_OP_IM2COL},
+    {VAD_TENSOR_ENC_0_BIAS,          GGML_OP_ADD},
+    {VAD_TENSOR_ENC_1_WEIGHT,        GGML_OP_IM2COL},
+    {VAD_TENSOR_ENC_1_BIAS,          GGML_OP_ADD},
+    {VAD_TENSOR_ENC_2_WEIGHT,        GGML_OP_IM2COL},
+    {VAD_TENSOR_ENC_2_BIAS,          GGML_OP_ADD},
+    {VAD_TENSOR_ENC_3_WEIGHT,        GGML_OP_IM2COL},
+    {VAD_TENSOR_ENC_3_BIAS,          GGML_OP_ADD},
+
+    {VAD_TENSOR_LSTM_WEIGHT_IH,      GGML_OP_MUL_MAT},
+    {VAD_TENSOR_LSTM_WEIGHT_HH,      GGML_OP_MUL_MAT},
+    {VAD_TENSOR_LSTM_BIAS_IH,        GGML_OP_ADD},
+    {VAD_TENSOR_LSTM_BIAS_HH,        GGML_OP_ADD},
+
+    {VAD_TENSOR_FINAL_CONV_WEIGHT,   GGML_OP_IM2COL},
+    {VAD_TENSOR_FINAL_CONV_BIAS,     GGML_OP_ADD}
+};
+
+static const std::map<vad_tensor, const char *> VAD_TENSOR_NAMES = {
+    {VAD_TENSOR_STFT_BASIS,          "_model.stft.forward_basis_buffer"},
+    {VAD_TENSOR_ENC_0_WEIGHT,        "_model.encoder.0.reparam_conv.weight"},
+    {VAD_TENSOR_ENC_0_BIAS,          "_model.encoder.0.reparam_conv.bias"},
+    {VAD_TENSOR_ENC_1_WEIGHT,        "_model.encoder.1.reparam_conv.weight"},
+    {VAD_TENSOR_ENC_1_BIAS,          "_model.encoder.1.reparam_conv.bias"},
+    {VAD_TENSOR_ENC_2_WEIGHT,        "_model.encoder.2.reparam_conv.weight"},
+    {VAD_TENSOR_ENC_2_BIAS,          "_model.encoder.2.reparam_conv.bias"},
+    {VAD_TENSOR_ENC_3_WEIGHT,        "_model.encoder.3.reparam_conv.weight"},
+    {VAD_TENSOR_ENC_3_BIAS,          "_model.encoder.3.reparam_conv.bias"},
+    {VAD_TENSOR_LSTM_WEIGHT_IH,      "_model.decoder.rnn.weight_ih"},
+    {VAD_TENSOR_LSTM_WEIGHT_HH,      "_model.decoder.rnn.weight_hh"},
+    {VAD_TENSOR_LSTM_BIAS_IH,        "_model.decoder.rnn.bias_ih"},
+    {VAD_TENSOR_LSTM_BIAS_HH,        "_model.decoder.rnn.bias_hh"},
+    {VAD_TENSOR_FINAL_CONV_WEIGHT,   "_model.decoder.decoder.2.weight"},
+    {VAD_TENSOR_FINAL_CONV_BIAS,     "_model.decoder.decoder.2.bias"}
+};